Our backtest data across the NMHC top 50 shows that bad debt reduction in recent years is largely due to macroeconomic improvement – not because tools like fake paystubs detectors solved the fraud problem. Despite marketing claims, these tools are not a holistic fraud prevention, screening, or consumer underwriting tools.
What is ‘document forensics’?
During Covid, affordability stress rose, and more applicants turned to fraud to get approved. In response, screening vendors began using a combination of machine learning and offshore review teams to detect fake PDFs.
The PDF file type was created in the 1990s to mimic security features contained in legal documents; they carry a unique identifier referred to as “metadata”, which makes it possible to securely link a document to its original source.
For example, a bank statement pulled straight from Bank of America contains metadata unique to the bank; edit one of these originals and that metadata no longer matches the institution’s.
While PDF metadata is a useful tool for verifying downloads from large institutions – as an anti-fraud product, it has its fair share of problems.
Problem 1: Verification requires original PDFs which creates friction
This technology relies on metadata that’s only contained in original PDFs. Photos, screenshots, or any other file type cannot be used to verify a document’s authenticity.
One issue is many applicants don’t submit PDFs at all. They submit photos of paystubs taken on their phone, screenshots from their banking app, or other file types. Photos, screenshots, and non-PDF file types all destroy metadata – making the submissions unverifiable.
The same is true when applicants combine PDFs or highlight direct deposits on bank statements – to make it easier on the reviewer. Or when they annotate documents for privacy – for example, blacking out sensitive info like account numbers. Every one of those destroys the metadata the system depends on.
These applicants get bounced back to manual review, and teams are left with the time-consuming task of reaching out and explaining that they have to log into their bank account, find the statements section, hit “download,” and resubmit the original file.
This creates an enormous amount of friction for applicants and teams. Some applicants may sign elsewhere, be unresponsive, reluctant to restart the process, or leave a frustrated review. While teams need time for responsibilities beyond repeatedly explaining why and how to resubmit.
Two Dots’ internal data shows 57% of applicants fail to upload income documentation correctly on their first attempt (even when given detailed instructions). With a document anti-fraud provider, every one of these applicants requires manual intervention and additional back-and-forth communication for resubmissions.
Problem 2: False positives are common
Combining PDFs to make reviewing easier, or annotating documents for privacy – like blacking out sensitive info such as account numbers – destroys original metadata the same way photos and screenshots do. In addition to manual review delays, a photo or modified PDF can wrongfully trigger the fraud detection system.
An altered field PDF is a common fraud indicator. The issue is this kind of system can’t see the intention behind an edit – harmlessly blacking out an account number and fraudulently photoshopping a balance end up with the same warning. The technology isn’t sophisticated enough to identify what the edit is, only that it’s been edited. False positives result from this inability to distinguish a harmless modification from a fraudulent one.
If the false positive contributes to a denial, operators are providing formal adverse action letters with decisions based on misinformation.
The risk of noncompliance carries statutory or actual damages per violation. This is what we see at scale in class action cases like the October 2025 Arimitsu v. AvalonBay/Snappt suit, which alleges systematic denial of rental applications on erroneous or arbitrary grounds.
Fraud tools are meant to reduce risk, not create it.
Problem 3: Many documents have no verifiable metadata
Some of the documents applicants submit don’t contain any verifiable metadata to begin with because they aren’t generated by institutions. Offer letters are the clearest example – roughly 10% of renters submit offer letters as proof of income.
Offer letters, real or fake, are mostly made in Microsoft Word or Google Docs by an individual. There isn’t a known ‘signature’ to verify against. The document forensics category was built to detect tampering on institutionally-generated documents, not personalized letters.
The verification work falls back on the leasing desk. Teams call employer phone numbers, send verification emails, try to confirm whether the person answering the phone is a real employee, and whether the company itself is real. The process is extremely time consuming, inconsistent – allowing fraud to slip by.
Applicants with offer letters are often some of the most qualified in the pool – people relocating for a new job, or people with a higher paying job looking to upgrade. These are also the applicants who don’t have extra days to wait for manual review with start dates approaching.
Problem 4: Models have to be trained for each institution
“Document forensics” works by training individual models for each institution – bank by bank, payroll provider by payroll provider – plus continuous upkeep as institutional formats constantly evolve. For vendors this consumes a significant amount of time and resources, so they prioritize the largest institutions where applicant volume and training data are highest. Smaller institutions get deprioritized.
Models can’t verify with metadata when they don’t know what to verify against, so applicant statements from smaller banks end up unverifiable at a higher rate. This affects members of community credit unions and lesser known banks, and employees of small businesses who likely use smaller platforms or do payroll manually.
Problem 5: Faking documents is just one type of fraud
Document anti-fraud is effective when conditions are met for its narrow use case.
However, document antifraud leaves portfolios exposed to many other fraud types, including: synthetic identity (or CPN) fraud, traditional identity theft, offer letter fraud, co-applicant / guarantor fraud, friendly fraud, and obfuscation fraud.
Falsifying income documents also has the lowest correlation with 60 day default using our backtest data from NMHC top 50 operators. Intuitively, you can imagine the average document fraudster as one who meets the credit requirements but just has low or non-traditional income and intends to pay the rent.
CPN and synthetic ID fraud on the other hand has some of the highest correlation with default, since the fraudsters are often hiding bad credit, which means they have not paid their obligations in the past. This is just one of the many types of fraud that get past these document based systems.
Finding the right solution:
These tools were marketed as holistic solutions for the multifamily bad debt problem. In reality, they prevented one narrow type of fraud while creating enormous friction for qualified applicants that hurt occupancy.
Meet Evie:
Evie is a conversational AI Approval Agent trusted by industry leaders – like BH Management and Weidner Apartment Homes.
Evie uses holistic underwriting to approve more applicants with less risk and optimize property performance:
- Prevents overrides and all fraud types – Evie prevents all fraud types and reaches decisions in 100% of applications. If someone submits an offer letter, Evie does online research and contacts employers to authenticate in real-time. No flags and no manual review where overrides flow through.
- Approves modern renters, with less risk – Evie verifies all income types and underwrites modern applicants with a special scoring model, allowing operators to optimize performance with property-level risk thresholds.
- Evie is proactive – Evie reminds applicants to complete the application, guides them through each step, answers questions, gathers additional documentation for complex cases – like thin or no-credit, and asks about guarantors or additional income sources before denial. Where others return false-positives or reasons to deny, Evie finds legitimate reasons to approve.
The first step is a backtest.
We run free backtests on actual portfolio data – every application from the past 12-24 months, screened against Evie, compared against what your existing process produced. The output shows where missed fraud is sitting in your current resident base, where overrides are concentrated, and what holistic underwriting would change.
The diagnosis in this report is general. The backtest is your portfolio’s specific version of it.
Appendix: Fraud Types
Income document fraud. Altered or fabricated paystubs, bank statements, tax documents, benefits letters. Was the most common fraud type in multifamily until recently being overtaken by synthetic identity fraud. Many applicants who commit this type of fraud do not default – they have passable credit, and are simply rent-burdened.
The correlation between this fraud type and 60-day default is roughly 30-40% during benign economic conditions; that correlation would likely rise during stress periods.
Synthetic identity fraud (CPNs). When an applicant uses a fabricated Social Security Number – a “Credit Privacy Number” or CPN – paired with their real name and identity. The CPN is associated with a clean (often thin) credit profile built specifically for the purpose. Now the most common fraud type in multifamily and the most highly correlated with default. CPNs grew from roughly 1.5% of rental applications in 2022 to over 4% in 2026 – a 260%+ increase. See §5 for the full mechanism.
Traditional identity theft. When an applicant pretends to be a different person entirely, using stolen SSN, photo ID, and other identity documents. Relatively rare in multifamily – under 0.5% of applications – because the fraudster has to take physical possession of the keys and unit, which creates exposure that other fraud types avoid.
Offer letter fraud. Fake offer letters submitted in lieu of paystubs to satisfy income requirements. Often paired with fake employer phone numbers (answered by accomplices) and fake company websites. Over 10% of offer letters Two Dots processes are fraudulent. The most-recommended workaround in online forums for beating standard screening, because most systems cannot meaningfully verify offer letters.
Co-applicant / guarantor fraud. Adding a fake co-applicant – typically with strong income and credit – to an application that would otherwise be denied. The co-applicant has no intention of actually living in or paying for the unit. Often appears suddenly when the primary applicant is about to be denied, after no co-applicant was mentioned in the original application.
Friendly fraud. When a friend or family member voluntarily lends their identity (SSN, photo ID, bank login) to someone with bad credit or a criminal history so that person can pass screening. Sometimes the lender applies and picks up the keys, after which someone else moves in. No fake documents and no synthetic IDs – just informal trust networks routing around screening criteria.
Inception fraud. When an applicant starts a fake small business and signs up for a real payroll provider like Gusto or QuickBooks. They then transfer money into a business bank account, pay themselves a direct deposit into their personal account, and immediately transfer the money back to the business bank account. This fraud type is relatively rare and is highly correlated with other fraud types like CPNs.
Obfuscation fraud. Deliberately withholding or misstating identity information (typically SSN) to prevent the screening system from accessing real credit, criminal, or rental history. Difficult to handle because legitimate applicants without SSNs (some recent immigrants) and fraudulent applicants hiding bad records produce identical workflow outputs: missing identifiers and incomplete data. New laws in Colorado and California prohibit landlords from asking about immigration status, which limits operators’ ability to distinguish the two populations.